LiteVLA-H, a 256M-parameter vision-language-action model, achieves dual-rate inference on NVIDIA Jetson AGX Orin for aerial drone guidance, issuing action tokens at 50.65 ms while supporting semantic outputs at 149.90–164.57 ms.

LiteVLA-H is a compact 256M-parameter VLA (vision-language-action) system—a model that combines visual perception, language understanding, and robotic control—designed to run on an NVIDIA Jetson AGX Orin embedded platform for onboard drone operation under strict compute and communication constraints.

The system operates in two concurrent modes: a fast outer-loop guidance mode that outputs action tokens at 50.65 ms (19.74 Hz), and a slower semantic mode for scene understanding and narration at 149.90–164.57 ms (6.08–6.67 Hz). This dual-rate design uses a scheduler that prioritizes multimodal pre-fill (the step of processing images and text context) as the latency bottleneck rather than token decoding.

The measured action branch achieves a higher edge inference rate than recent architectures including AnywhereVLA, FutureVLA, and ReMem-VLA under the same deployment conditions while retaining periodic semantic awareness.